Research On Image Super-resolution Reconstruction Algorithm Based On Wavelet Transform And Self-Attention

Recently,deep convolution neural networks(DCNN)have been widely explored in single image super-resolution(SISR)and obtained remarkable performance.However,most of the existing CNN-based SISR methods tend to produce over-smoothed outputs,and the recovery of image details is not ideal enough.To address these issues,this paper conducts related research on image super-resolution reconstruction algorithm based on wavelet transform and self-attention mechanism,and proposes two image super-resolution reconstruction algorithms:(1)Wavelet-based Asymmetric Convolution Network(WACN).The existing super-resolution methods do not make full use of the high-frequency information of the image,resulting in unclear details of the reconstructed image.WACN reconstructs the wavelet coefficients in high-resolution images by linking the super-resolution magnification and the wavelet decomposition series,and the network can focus on learning the high-frequency information of the image by combined image space loss and wavelet coefficient loss.In addition,WACN reconstructs the residual block and designed an asymmetric convolution residual block.In the training phase,asymmetric convolution residual block can provide different receptive fields to enrich feature information.In the inference phase,asymmetric convolution kernel can be equivalently fused into the standard square-kernel layers,such that no extra computational burdens are introduced in the inference phase,improving the speed of reasoning.(2)Combining Attention in Wavelet-based Asymmetric Convolution Network(CAWACN).In order to make the deep super-resolution network be able to adaptively distinguish high and low frequency information,and further improve the utilization of high frequency information,this paper proposes a channel attention module based on variance(VCA),and connects it to WACN to get CAWACN.VCA mechanism can adaptively rescale channel-wise features by considering interdependencies among channels,which improved reconstruction effectiveness.The final experimental results show that these two methods can make full use of the high-frequency information of the image,so that the high-frequency edge detail information of the reconstructed image is rich,and it is closer to the real high-resolution image.